Predicting the status of COVID-19 active cases using a neural network time series

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 3, June 2022, pp. 3104~3117
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i3.pp3104-3117  3104
Journal homepage: http://ijece.iaescore.com
Predicting the status of COVID-19 active cases using a neural
network time series
Peyman Almasinejad1
, Amin Golabpour2
, Fatemeh Ahouz3
, Mohammad Reza Mollakhalili Meybodi1
,
Kamal Mirzaie1
, Ahmad Khosravi4
, Marzieh Rohani-Rasaf5
, Azadeh Bastani3
1
Department of Computer Engineering, Maybod Branch, Islamic Azad University, Maybod, Iran
2
Department of Health Information Technology, School of Allied Medical Sciences, Shahroud University of Medical Sciences,
Shahroud, Iran
3
Department of Computer Engineering, School of Engineering, Behbahan Khatam Alanbia University of Technology, Behbahan, Iran
4
Ophthalmic Epidemiology Research Center, Shahroud University of Medical Sciences, Shahroud, Iran
5
Department of Epidemiology, School of Public Health, Shahroud University of Medical Sciences, Shahroud, Iran
Article Info ABSTRACT
Article history:
Received May 17, 2021
Revised Dec 29, 2021
Accepted Jan 12, 2022
The design of intelligent systems for analyzing information and predicting
the epidemiological trends of the disease is rapidly expanding because of the
coronavirus disease (COVID-19) pandemic. The COVID-19 datasets
provided by Johns Hopkins University were included in the analysis. This
dataset contains some missing data that is imputed using the multi-objective
particle swarm optimization method. A time series model based on nonlinear
autoregressive exogenou (NARX) neural network is proposed to predict the
recovered and death COVID-19 cases. This model is trained and evaluated
for two modes: predicting the situation of the affected areas for the next day
and the next month. After training the model based on the data from January
22 to February 27, 2020, the performance of the proposed model was
evaluated in predicting the situation of the areas in the coming two weeks.
The error rate was less than 5%. The prediction of the proposed model for
April 9, 2020, was compared with the actual data for that day. The absolute
percentage error (AE) worldwide was 12%. The lowest mean absolute error
(MAE) of the model was for South America and Australia with 3 and 3.3,
respectively. In this paper, we have shown that geographical areas with
mortality and recovery of COVID-19 cases can be predicted using a neural
network-based model.
Keywords:
Epidemic outbreak
Missing value
Neural network
Time series prediction
This is an open access article under the CC BY-SA license.
Corresponding Author:
Amin Golabpour
Department of Health Information Technology, School of Allied Medical Sciences, Shahroud University of
Medical Sciences
Shahroud University of Medical Sciences and Health Services, Hafte Tir Square, Shahrood, Iran
Email: a.golabpour@shmu.ac.ir
1. INTRODUCTION
On December 8, 2019, the Chinese government reported the death of one patient and the
hospitalization of 41 others with an unknown etiology in Wuhan [1], [2]. This cluster initiated the novel
coronavirus disease (COVID-19) respiratory disease pandemic. While early cases of the disease were linked
to the wet market, the human-to-human transmission had led to the widespread outbreak of the virus through
China [3]. On January 30, 2020, the World Health Organization (WHO) announced the emergence of
COVID-19 as a public health emergency with international concern (PHEIC) [4].
By March 16, 2020, WHO had reported the COVID-19 statistics in China and outside of China.
However, since March 17, 2020, because of widespread prevalence on all continents, the number of

Int J Elec & Comp Eng ISSN: 2088-8708 
Predicting the status of COVID-19 active cases using a neural network time series (Peyman Almasinejad)
3105
confirmed cases and deaths on each continent was separately expressed [4]. According to the 49th
COVID-19
weekly epidemiological update by WHO released on 20 July 2021, globally, COVID-19 weekly case
incidence increased with an average of around 490,000 cases reported each day. As of 18 July 2021, the
global number of confirmed cases was 190,169,833 [5].
Because of the widespread and growing prevalence of COVID-19 across the world, several works
have examined different aspects of the disease. Most of these include identifying: i) the source of the virus
and its gene sequences analysis [6], [7], ii) analysis of patient information [8], iii) analysis of the first cases in
the countries involved [9]–[11], iv) methods of virus detection [12]–[15], v) evaluation of treatment methods
[16], and vi) estimating the extent of transmission [17]. Artificial intelligence has an important role in
changing the medical care paradigm and can predict various diseases states [18]–[23]. Thus, scores of
research have been done in the past year in areas related to the COVID-19 pandemic. The most common
topics are: i) the development of health care robots to prevent direct contact of medical personnel with
COVID-19 patients [24], [25]; ii) monitoring of public places to determine the distance between persons or
identification of people with high temperature [26]–[28]; iii) forecasting the spread of COVID-19 [29]–[31];
iv) automatic diagnosis of patients with COVID-19 [32]; and v) predicting confirmed, recovered, or death
cases [33], [34].
Fang et al. [35] proposed a methodology called group of optimized and multisource selection
(GROOMS), which is an ensemble of 5 groups of prediction methods. They also proposed a new version of
polynomial neural network (PNN) called “PNN with corrective feedback (PNN+cf)” that includes two extra
pieces of information: i) lagged data and ii) training errors from past iterations of model training to predict
the epidemic at an early stage. The authors conducted an experiment on epidemic data from Chinese health
authorities from January 21 to February 3 to evaluate the initial stage of the COVID-19 epidemic. A time-
series of 14 instances about the suspected cases was run through the GROOMS method for 6-days ahead of
the forecast. They compared the result of the model to the other nine available methods and claimed that the
PNN+cf method with 136,547 root-mean-square-error (RMSE) was better than the other methods. The
RMSE of the other methods was reported from 138.042 to 1744.5256.
Corona tracker team proposed a susceptible-exposed-infectious-recovered (SEIR) model based on
the queried data in their website and made the 240-day prediction of COVID-19 cases in and out of China,
started on 20 January 2020 [36]. They predicted that the outbreak would reach its peak on May 23, 2020, and
the maximum number of infected individuals will be 425.066 million globally. The authors predicted that it
would start to drop around early July 2020 and reach under 10,000 on 14 Sep 2020. Given the information
available now, these predictions were far from what really happened around the world.
Wang et al. [37] constructed a COVID-19 prediction model using the improved long short-term
memory (LSTM) deep learning method with a rolling update mechanism based on the epidemical data
provided by Johns Hopkins University. The trends of the epidemic in 150 days ahead were modeled for
Russia, Peru, and Iran. Pointing to the importance of preventive measures which would be taken by the
government to reduce the spread of COVID-19, the authors estimated that the number of positive cases per
day in Iran by mid-November 2020 will reach less than 1,000, however, it did not happen.
Zawbba et al. [38] proposed a regression model based on the multilayer perceptron (MLP) to predict
the COVID-19 spread for the coming months in nine countries: Italy, the United States, China, Japan, Iran,
Egypt, Algeria, Kenya, and Cote d'Ivoire. For each country, they used the number of confirmed cases, the
number of deaths, average age, average weather temperature, Bacille Calmette-Guérin (BCG) vaccination,
and Malaria treatment. The model was first trained on Chinese data, which were collected from the European
Centre for Disease Prevention and Control (ECDC) from 29th
December 2019 to 13th
December 2020. Then,
for the other eight countries, data were downloaded from 22nd
January 2020 to 13th
December 2020 from the
Johns Hopkins University Center for Systems Science and Engineering (JHU CSSE). The best and worst
RMSE for confirmed cases were 105.94 and 77,822.38 in Cote d'Ivoire and the United States, respectively. In
the case of predicting the dead cases, the best RMSE was reported 0.91 in Cote d'Ivoire, and the worst one
was 792.07 in the USA.
Ahouz and Golabpour [29] introduced a new representation structure of the COVID-19 dataset. By
dividing the data set regions into three groups based on the maximum number of confirmed cases of
COVID-19 per day and using the least-squares classification algorithm, the authors developed models for
predicting the incidence of COVID-19 between March 30, 2020, and April 12, 2020, for each group. The
accuracy of the model in predicting the number of COVID-19-approved cases worldwide is reported to be
98.45%.
A review of COVID-19 research shows that most of the predictive models utilized several months
of COVID-19 information [20], [34], [38], [39]. However, at the very beginning of pandemics such as
COVID-19, due to the lack of information about the factors affecting the recovery or death of patients, design
models with appropriate accuracy based on limited non-clinical information is very important. These models

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 3, June 2022: 3104-3117
3106
are useful in the containment of the threat and help healthcare administrators to make effective timely
decisions in controlling the spread of the disease in addition to reducing community anxiety.
In this study, using less than two-month prevalence data, we proposed a neural network-based time
series model for predicting the recovery or death of COVID-19 patients based on general information of each
region including latitude, longitude, date, and number of confirmed/recovered/death cases. Based on the
datasets provided by Johns Hopkins University [40], we present a new arrangement of the data for the
optimal use of outbreak trend information in neighboring regions. In addition, using the information from
each region, we predict the final status of COVID-19 cases in each region in the next day and month.
2. METHOD
2.1. Dataset
COVID-19 epidemiological data are compiled by the Johns Hopkins University Center for Systems
Science and Engineering (JHU CCSE) [31]. The data are provided in three separate datasets for confirmed,
recovered, and death cases since January 22, 2020, and are updated daily. In each of these datasets, there is a
record (row) for each geographic area. The variables in each dataset are province/state, country/region,
latitude, longitude, and then incremental dates from January 22, 2020. For each geographic area, the value of
each date indicates the cumulative number of confirmed/recovered/death cases from January 22, 2020.
In this study, data from the COVID-19 dataset from January 22 to March 9, 2020, entered into the
analysis. This information includes the number of confirmed, recovered, and death cases in 265 different
geographical areas in 47 days. According to the input requirements of the proposed model, we changed the
data representation in the dataset so that instead of three separate datasets for the three groups of confirmed,
recovered, and death cases, only one dataset containing the information of all three groups was arranged. In
this new dataset, each record (or row) of the dataset contains information about the number of confirmed,
recovered, or deaths per day for each geographic area. As a result, the variables in this new dataset are
province/state, country/region, latitude, longitude, date (which specifies a specific date), cases (which
indicates the number of confirmed, recovered, or death cases), and type (which specifies the type of
confirmed, recovered, or death cases). This structure was suggested by Krispin [41].
This rearranged dataset contains 3,436 records and 7 variables that include information on
COVID-19 cases from 111 countries and 265 different geographic regions around the world. There are
113,583 confirmed, 3,996 death, and 62,512 recovered cases in the dataset. Pre-processing runs on the
dataset before training the proposed model. The dataset is first examined for noise data. Then, the missing
data were investigated, and it was found that the data were recorded with a delay, and 56 records out of 256
were missed. To impute the missing data the following procedure was used:
First, the variable is sorted according to the amount of missing data. Put the variables that have the
least missing at the beginning, and the variables that have the most missing at the end. Then, with the help of
data mining algorithms, a classification model is generated whose independent variables are the variables that
do not exist in the missing data, and its dependent variable is a variable with the lowest rate of missing. Then,
with the help of the data mining algorithm, the missing data of the dependent variable are imputed. The
dependent variable is then added to the set of independent variables, and the next variable is selected as the
dependent variable according to the number of missing data, and the data mining algorithm is applied again.
This process continues until there is no variable with the missing data. The selection of data algorithms is
performed by the multi-objective particle swarm optimization algorithm. Repeat the above process several
times so that no change is made to the data. At the end of this process, all the data are filled in, and there are
no missing data.
Then, depending on the need of the learning algorithm used in the proposed model, the values of
some variables are changed to another format. If necessary, some variables are merged, and new variables are
added to the dataset. There are 24 negative values in the cases which are invalid, so we removed them from
the dataset. As a result, the number of records decreased to 3,412.
In the province/state, there were 901 missing data out of 3,436 because the information of some
countries was generally reported in that country and not for a particular province. Therefore, we imputed the
missing data with the name provided in the country region column. We then assigned a unique numeric code
to each of these 265 regions. This new code is called code_zone and will be added to the dataset instead of
the country and state columns. After applying these preprocessing steps, the resulting dataset is called the
alpha dataset.
The records in the alpha dataset were sorted by the value of the date variable. Since in time series
models, the algorithm detects the time series pattern, the date variable is removed from the dataset. This leads
to the construction of the beta dataset. Therefore, after preprocessing steps, the beta dataset includes 3,412
records with 5 variables (code_zone, latitude, longitude, cases, and type).

3107
Finally, the beta dataset is compiled for two separate two-class prediction models, namely recovery
prediction and death prediction. To predict recovered cases, records with "recovered" value in the type
variable are considered as class 1, and other records are considered class 0. To predict death cases, records
with "death" value in type are considered as class 1, and other records are considered as class 0.
Descriptive statistics such as mean and standard deviation were used to describe the variables.
Kruskal Wallis test was used to compare the mean of longitude and latitude variables in the confirmed, death,
and recovered groups. This test is to examine the significance of the relationship between latitude and
longitude with mortality and recovery. If the test is approved, the latitude and longitude can be used as
metadata in the model.
2.2. Constructing the time series prediction model
Because of the presence of metadata in the dataset, an algorithm that utilizes this information to
construct time series is preferable. The nonlinear autoregressive exogenous (NARX) neural network is one of
these methods [42]. The output of the NARX network at time t+1 is given using (1):
𝑦(𝑡) = 𝐹(𝑦(𝑡 − 1), 𝑦(𝑡 − 2), . . . , 𝑦(𝑡 − 𝑛𝑦), 𝑥(𝑡), 𝑥(𝑡 − 1), 𝑥(𝑡 − 2), . . . , 𝑥(𝑡 − 𝑛𝑦)) (1)
where 𝐹(. ) is the mapping function of the neural network, 𝑦(𝑡 − 1), 𝑦(𝑡 − 2), . . . , 𝑦(𝑡 − 𝑛𝑦) are the true past
outputs of this series, called the desired outputs. And 𝑥(𝑡), 𝑥(𝑡 − 1), 𝑥(𝑡 − 2), . . . , 𝑥(𝑡 − 𝑛𝑦) are the inputs of
the NARX which are called exogenous inputs; these metadata are externally determined and influence the
desired output of the series. nx is the number of input delays, and ny is the number of output delays [42]. In
the proposed model, we aim to determine which areas will experience the death or recovery of COVID-19
cases in the next day. This model is general and does not depend on a specific area. For this reason, we do
not use the present value of 𝑥(𝑡). Thus, the future value of the time series 𝑦(𝑡) is predicted from the past
values of 𝑥(𝑡) and the actual past values of the time series, 𝑦(𝑡).
Figure 1 illustrates the preparation steps of the NARX inputs. In Figure 1, n represents the number
of records, 𝑦
̂ and type are the predicted and actual outputs, respectively, which in the death prediction model
can be dead (1) or not-dead (0), and in the recovery prediction model can be recovered (1) or not-recovered
(0). In addition, latitude, longitude, cases, and their types on day t-d to day t-1 are the exogenous inputs of the
network. Given n record information up to time t-1, the proposed model attempts to predict whether
COVID-19 active cases will recover (die) at time t in a geographic area. It should be noted that for each
geographic area in the dataset, the past information of that area is used to predict its 𝑦
̂(𝑡). In the model, input
and output delays are represented by d or the delay factor.
Figure 1. The preparation steps of the NARX input

 ISSN: 2088-8708
3108
The delay factor determines how many past observed data of a particular area will participate in
predicting the one-step-ahead situation of that area. If the delay is one, it means the model uses only one
previously observed point to predict the type of cases at time t. Similarly, if it is k, by using k past observed
data points, the model predicts the future at time t. The larger the k, the more information is used to build the
model, thus increasing the accuracy of the model as well as its complexity.
In neural network-based time series models, inputs and structure of networks are of great
importance. The most important factors are: i) determining the number of neurons in the hidden layers, ii) the
learning algorithm for adjusting the weights of the network, and iii) determining the amount of past-observed
data to predict the future. In this study, to avoid over-complexity of the network, the maximum number of
neurons in hidden layers and the maximum delay are predetermined. For each learning algorithm to adjust
network weights, the number of different neurons in the hidden layer is evaluated. For each different number
of hidden neurons, different values of d are examined, and the network is trained on the training set. After
training, the performance of the network is evaluated using an evaluation dataset, and the weights of the
network are updated. Figure 2 shows the steps of the proposed model.
Because the latitude and longitude of each geographic area are entered as metadata in the model, the
weights of the neural network are affected by the coordinates of all geographic areas. This allows us to train
just one neural network for all areas in the dataset, instead of training one network for each geographic area.
After completing the training and evaluating the model, it is even possible to enter the information of new
areas into the network and predict the death or recovery status of those areas without the need for retraining.
The final time series prediction model is evaluated on the test set, and the performance of the model is
evaluated based on sensitivity, specificity, and accuracy. Finally, considering all combinations of hidden
neurons and different delays, the best predictive model is reported.
Figure 2. The steps of the proposed method
2.3. Predicting the status of active cases
Active cases of COVID-19 are those who have not yet died or recovered. Therefore, the difference
between the total numbers of confirmed cases and the total number of dead or recovered cases in the Beta

3109
dataset are considered active cases. After finding the best models for predicting death and recovery, these
models will be used to predict the status of active cases.
To do this, a new test set containing 265 records will be created for each of the 265 different regions
in the dataset. For each record, information about code_zone, Latitude, Longitude, and the number of active
cases is provided. In addition, we will add date information to this new dataset to predict the situation of each
region in the next month. The value of the date variable for all 265 geographic areas is set to one month after
the last date in the Alpha dataset. Then, the best proposed models for predicting death and recovery will be
trained using the Alpha dataset, which contains the date variable. At this point, the entire dataset is used for
training and evaluation. After training, the models are run on the new test dataset. If both death and recovery
models are assigned a record to the negative class (class label 0), the record type is set to be active.
Otherwise, the type is defined depending on the output of the models, i.e. recovered, death, or both. The
model predicts what the status of active cases will be in each geographic area in the next month. Depending
on the number of areas in which patients' status is active, dead, or recovered, the number of areas in which
confirmed cases of COVID-19 will be expected is calculated according to (2):
𝐶𝑜𝑛𝑓𝑖𝑟𝑚𝑒𝑑 = 𝐴𝑐𝑡𝑖𝑣𝑒 + 𝐷𝑒𝑎𝑡ℎ + 𝑅𝑒𝑐𝑜𝑣𝑒𝑟𝑒𝑑 (2)
Finally, to see how well the model predicts areas with COVID-19 death, improved, or confirmed
worldwide in the next month, the model's prediction results are compared with the actual data. The criteria
are absolute error (AE), mean absolute error (MAE), and absolute percentage error (APE). For each type of
case, the percentage of its occurrence is obtained by dividing the number of areas of that type by the sum of
the total number of areas of all three types.
The experimentation platform is Intel® Core ™ i7-8550U CPU @ 1.80GHz 1.99 GHz CPU and
12.0 GB of RAM running 64-bits OS of MS Windows 10. The SPSS version 15 was used for descriptive and
statistical analysis. The pre-processing and model construction have been implemented in MATLAB.
3. RESULTS
3.1. Descriptive analysis
Data as of March 9, 2020, showed 113,583 confirmed cases, 3,996 deaths, and 62,512 recovered
cases. Figure 3 shows the distribution of confirmed, recovered, and death cases by latitude and longitude. The
mean distance from the equator in all three groups was significant. The confirmed cases were farther from the
equator than other groups, and those recovered were closer to the equator (P<0.001). The highest mean
distance from prime meridian was related to the recovered group (P<0.001). Kruskal Wallis test showed that
all groups were significantly different in latitude, and the recovered group was significantly different from
other groups in longitude as shown in Table 1.
Figure 3. Distribution of recovered, death, and confirmed cases of coronavirus dataset over the world

 ISSN: 2088-8708
3110
Table 1. Descriptive analysis of COVID-19 dataset [31] using Kruskal Wallis test
Attribute name Confirmed Death Recovered Chi-Square SIG
Mean±SD Mean±SD Mean±SD
Latitude 33.30±05.89 32.88±04.51 31.19±03.28 30.633 P<0.001
Longitude 96.24±37.47 95.88±36.05 109.18±16.51 175.627 P<0.001
3.2. Model construction
In this study, scaled conjugate gradient backpropagation (SCG) [43] and Levenberg-Marquardt
backpropagation (LM) [44] algorithms were used to train and adjust the weights of the NARX neural
network. With a step length of 5, the number of neurons in the hidden layer and the range of delay are varied
from 5 to 25 and 5 to 50, respectively. Each model for death and recovery prediction was trained separately
on the relevant dataset (interval 5 is obtained by trial and error) [42]. Each dataset is divided into 55%
training set, 15% evaluation set, and 30% test set. Since the dataset has a natural temporal order, these values
are selected in the dataset with the same order. Thus, the training dataset contains the first 55% of the
records, the evaluation set contains 15% of the next records, and the test set contains 30% of the last data
records. As a result, data from January 22 to February 27, 2020, were used to build the model
(training+evaluation datasets), and data from February 28 to March 9, 2020, were used to test the model. To
do this, the "divideblock" function is selected as the neural network
Table 2 and Table 3 show the results of the proposed LM-based model on the test set to predict
recovery and death for all combinations of hidden layer neurons and delay factors, respectively. The results
are accepted or rejected based on two criteria: the result is discarded: i) if the performance of the model on
the test set is better than that on the training set or ii) the model accuracy is less than 0.75%. Accordingly,
values written in gray in the tables indicate unacceptable values. Given these two criteria, there is no valid
case among the different modes of the proposed SCG-based model for predicting recovered cases. However,
there is a single valid mode among the different SCG-based models for predicting deaths. This model is built
of 25 nodes in the hidden layer and information from 5 previous records (delay). The accuracy, sensitivity,
and specificity were 96.27%, 73.5%, and 98.39%, respectively.
Table 2. The result of the proposed method for prediction of recovered cases using LM
Learning algorithms Number of past information
5 10 15 20 25 30 35 40 45 50
Levenberg-Marquardt backpropagation number of hidden neurons
5 Sensitivity Training 95.54 97.17 92.61 96.96 96.20 95.87 95.54 94.56 96.19 97.17
Test 96.74 96.38 96.38 96.38 96.38 96.01 95.65 95.65 95.65 95.27
Specificity Training 97.68 97.21 97.61 98.84 97.27 98.30 98.57 97.55 97.07 96.39
Test 92.07 98.24 98.64 97.39 98.48 97.91 84.17 79.41 87.78 64.29
Accuracy Training 96.86 97.19 95.68 98.11 96.86 97.36 97.40 96.40 96.73 96.69
Test 93.33 97.73 98.02 97.11 97.90 97.39 87.37 83.96 90.00 73.03
Test 96.74 96.38 96.38 96.74 96.74 96.01 95.65 95.65 94.93 95.64
Test 98.92 98.51 98.91 89.03 86.60 92.21 95.52 75.18 72.30 55.14
Test 98.33 97.93 98.22 91.14 89.40 93.27 95.56 80.91 78.67 66.56
Test 96.74 96.38 96.74 96.74 96.38 96.01 95.29 94.93 94.93 95.64
Test 98.52 98.78 87.74 87.24 99.03 80.25 98.32 95.63 98.58 98.43
Test 98.04 98.13 90.20 89.85 98.30 84.62 97.47 95.43 97.55 97.64
Test 96.74 96.38 96.74 96.38 96.38 95.65 95.65 95.29 94.20 95.27
Test 93.68 98.38 81.06 99.31 97.93 95.41 82.07 77.43 95.88 90.14
Test 94.51 97.83 85.35 98.51 97.50 95.48 85.86 82.44 95.41 91.59
Test 96.74 96.38 96.74 96.38 96.38 96.38 95.29 94.57 94.57 94.55
Test 98.39 98.51 89.92 96.43 97.79 80.95 80.53 97.32 83.66 89.29
Test 97.94 97.93 91.78 96.42 97.40 85.23 84.65 96.55 86.73 90.77

3111
Table 3. The result of the proposed method for prediction of death cases using LM
Learning algorithms Number of past information
5 10 15 20 25 30 35 40 45 50
Levenberg-Marquardt backpropagation number of hidden neurons
Test 75.86 85.06 47.13 59.77 71.26 50.57 64.37 67.82 79.52 54.43
Test 98.50 99.35 99.24 98.04 99.23 99.23 99.11 99.44 98.66 97.88
Test 96.57 98.13 94.75 94.73 96.80 94.97 96.06 96.65 97.04 94.36
Test 81.61 52.87 42.53 62.07 48.28 65.52 47.13 68.97 54.22 69.62
Test 98.93 99.14 98.27 99.02 98.80 97.80 94.02 99.44 99.67 97.21
Test 97.45 95.17 93.47 95.82 94.40 94.97 89.90 96.75 95.82 94.97
Test 60.92 64.37 44.83 48.28 52.87 40.23 80.46 85.06 53.01 53.16
Test 99.68 99.57 98.81 95.53 99.01 99.01 97.67 96.10 98.66 99.22
Test 96.37 96.55 94.16 91.44 95.00 93.87 96.16 95.13 94.80 95.49
Test 80.46 63.22 55.17 65.52 43.68 44.83 62.07 51.72 73.49 75.95
Test 98.93 98.71 98.70 97.17 99.67 91.96 97.45 95.43 99.11 96.88
Test 97.35 95.67 94.95 94.43 94.80 87.84 94.34 91.57 96.94 95.18
Test 78.16 50.57 72.41 59.77 65.52 58.62 52.87 71.26 45.78 51.90
Test 98.61 85.13 98.59 97.71 88.61 99.34 91.14 93.54 98.44 92.30
Test 96.86 82.17 96.34 94.43 86.60 95.78 87.78 91.57 93.98 89.03
To predict the recovered cases, 21 out of 50 possible combinations were accepted among different
modes of implementation of the proposed method using the LM learning algorithm. The best combination
consists of 15 neurons in the hidden layer and uses the last 25 observations. The accuracy, sensitivity, and
specificity of evaluation of this model on the test set were 98.30%, 96.38%, and 99.03%, respectively.
Also, to predict death cases, 27 out of 50 possible combinations were accepted among different
modes of implementation of the proposed method using the LM learning algorithm. The best combination
consists of 15 neurons in the hidden layer and uses the last 40 observations. Accuracy, sensitivity, and
specificity of evaluation of this model in the test set were 95.13%, 85.06%, and 96.10%, respectively.
The figure shows the best period until the neural network reaches the desired level. At this stage, the
mean square error for predicting recovered cases using the LM learning method is 0.026. To predict the
deaths, this value is 0.048 and 0.037 for LM and SCG learning methods, respectively. Henceforth, the model
behavior and error rate remain almost constant and the zigzag mode is not observed in the plot.
3.3. Predicting the status of the active patients
Because the best predictive model of the recovered cases was created using the LM learning
algorithm, consisting of 15 hidden layer neurons and 25-time delays, these settings were used to construct a
predictive model of the probability of recovery of active cases. The LM-based death prediction model
settings consist of 15 hidden layers and 40-time delays and are more sensitive. Thus, this setting was used to
construct the predictive model of the probability of death of active cases. The whole Alpha dataset (includes
information from January 22 to March 9, 2020) was used to train and evaluate each model with an 85%-15%
training-validation partition. The trained model was then run on the new test set containing 265 records. By
adding the Date variable to 265 records and setting it to the next month, in this case, April 9, 2020, we can
predict the status of active patients in the coming month. The results of the continental implementation are
presented in Table 4. For each type of case on each continent, the number of actual and predicted COVID-19
cases and the absolute errors of these two values are reported. The last column shows the MAE of the
proposed model in predicting the type of COVID-19 cases on each continent. Accordingly, the lowest MAE
of the model is obtained for South America and Australia with 3 and 3.33, respectively. The highest MAE is
for North America with 74.67. In predicting the number of areas that will have new confirmed cases of
COVID-19on April 9, 2020, the lowest MAE is for Europe with 1 area and the worst for North America with
126 regions. In predicting the number of regions leading to death, the best predictions are for Africa and Asia

 ISSN: 2088-8708
3112
with a MAE of 2 and the worst case for Europe with a MAE of 28. To predict the areas where people will
recover, the best prediction is for Australia with one area difference and the worst for North America with 79
areas. Globally, the best outcome among the predictions of areas with confirmed, recovery, and death is
related to the prediction of death with 13 cases.
Table 4. Prediction of the state of active COVID-19 patients by the proposed model
Row Continent Confirmed Recovered Death MAE
Predicted Actual AE Predicted Actual AE Predicted Actual AE
1 Africa 13.00 26.00 13 6.00 15.00 9 5.00 7.00 2 8.00
2 Asia 86.00 41.00 45 54.00 40.00 14 17.00 19.00 2 20.33
3 Australian 13.00 8.00 5 8.00 7.00 1 5.00 1.00 4 3.33
4 Europe 56.00 55.00 1 34.00 39.00 5 11.00 39.00 28 11.33
5 North America 150.00 24.00 126 87.00 8.00 79 33.00 14.00 19 74.67
6 South America 11.00 11.00 0 5.00 10.00 5 3.00 7.00 4 3.00
World 329 165 164 194 119 75 74 87 13 84.00
Figure 4 shows the actual and predicted percentage of the areas that will encounter the occurrence of
each of the three types of death, recovered, and confirmed cases on April 9, 2020. This figure shows the
percentage of occurrence of each type on each continent. In predicting areas with confirmed cases, the best
accuracy was obtained in Africa and Australia with 100%. In predicting areas with death cases, the best
accuracy was obtained in Africa with 93.75%, and the best recovery prediction was achieved for Europe with
95.66%. In addition, Figure 5 shows the distribution of different predicted types of COVID-19 cases in the
next month across the world.
Figure 4. The actual and predicted percentage of occurrences of confirmed, recovered, and death cases
around the world on April 9, 2020

3113
Figure 5. Distribution of death, recovery, and active COVID-19 infected patients across the world
4. DISCUSSION
In this study, the COVID-19 dataset derived from the JHU CSSE datasets of COVID-19, from
January 22 to March 9, 2020, was used. This dataset contains 3,436 records from 265 different geographic
regions. Each record stores the information about code_zone, latitude, longitude, cases, and type of
COVID-19 cases. The significance of latitude and longitude with type (confirmed-death-recovered) variable
with 99% confidence interval were evaluated by the Kruskal Wallis test. There was a significant relationship
between latitude and longitude and the status of covid-19 patients (P<0.001). Therefore, we entered these
features into the prediction models.
Neural networks are known as an accurate and powerful tool for solving complex and nonlinear
problems and predictive models [45]. Thus, a time-series prediction model was designed using a neural
network algorithm. This is an efficient method for predicting the status of confirmed, death, and recovered
cases of COVID-19.
After training the models on data from January 22 to February 27, 2020, the models were tested on
data from February 28 to March 9, 2020. The sensitivity, specificity, and accuracy of the model on the test
set were 96.38%, 99.03%, and 98.3%, respectively, to predict the recovery. The sensitivity was 85.06%,
specificity was 96.1%, and accuracy was 95.13% for mortality prediction. These results indicate that this
model is very suitable for predicting COVID-19 status.
The strength of the proposed model is that it does not require retraining to predict mortality or
recovery of affected areas in the short term. Since the training of neural network models is time-consuming,
eliminating the training time allows us to predict the situation of the next day of the regions with high speed
and accuracy by providing daily information. In this study, we evaluated the proposed model as a pilot to
predict the state of the regions for 12 consecutive days. The accuracy was higher than 95% in both models of
death and recovery prediction. Daily prediction of death and recovery status of affected areas is important in
the management of medical staff and timely measures.
In the second part of the prediction, information about one month after the selected period, i.e. April
9, 2020, for each continent, is expressed in terms of the number and percentage of infected areas. The results
of the proposed model were compared with the actual data in the updated JHU CSSE datasets of COVID-19.
In Africa, for all three types of recovered, deaths, and confirmed cases, the proposed method had maximum
and minimum prediction error of 13 and 2 geographical areas, respectively. In Asia, for all three types of
recovered, deaths, and confirmed cases, the proposed method had maximum and minimum prediction errors
of 45 and 2, respectively. In Australia, for all three types of recovered, deaths, and confirmed cases, the
proposed method has maximum and minimum prediction errors of 5 and 1, respectively. In Europe, for all
three types of recovered, deaths, and confirmed cases, the proposed method had maximum and minimum
prediction errors of 28 and 1, respectively. In North America, for all three types of recovered, deaths, and

 ISSN: 2088-8708
3114
confirmed cases, the proposed method had maximum and minimum prediction errors of 126 and 19,
respectively. In South America, for all three types of recovered, deaths, and confirmed cases, the proposed
method had maximum and minimum prediction errors of 5 and zero, respectively. The mean absolute
percentage error of all three types is 4.17% in Africa, 9.19 % in Asia, 8.65% in Australia, 12.29% in Europe,
12.14% in North America, and 12.41% in South America. According to Figure 4, after aggregating the
results of all continents, for all three types of confirmed, death, and recovered in the world, the proposed
model had a maximum prediction error of 164 geographical areas and a minimum of 13 geographical areas.
The maximum absolute percentage error for the world is 11.05%.
Another limitation of this study is the use of data from all countries involved in COVID-19, while
each country has its own protocol for testing and identifying patients. However, in general, this is the only
global dataset for COVID-19 that has been used in other studies [17], [29], [34], [46]–[48]. Also, in the
proposed model, the past information of each country has been used to predict the COVID-19 status of that
country, and this reduces the mentioned limitation. It is suggested that if the area is specifically defined,
variables such as temperature and humidity, weather conditions, and population density of the area should be
used in creating the model.
5. A DATA SHARING STATEMENT
Dataset is public and available on Johns Hopkins University Center for Systems Science and
Engineering (JHU CCSE). Novel coronavirus (COVID-19) cases [49].
6. CONCLUSION
COVID-19 has been dramatically spreading around the world during an epidemic, the speed of
information gathering and information dissemination is crucial to the containment of the threat. The mortality
or recovery of patients in a region/country or continent should be predicted because it helps with timely
action and medical decisions. In addition, resource management will be more effective if that the trend of
mortality or recovery in an area is predicted in the next week or two. Since epidemiological models such as
SIR are not able to accurately predict mortality and recovery of COVID-19 cases, we presented a more
complex model based on machine learning methods using the COVID-19 Cases dataset provided by Johns
Hopkins University. Therefore, due to the time-series nature of COVID-19 data, a neural network-based
time-series method is presented to predict the mortality or recovery status of COVID-19 cases in different
geographical areas around the world. In addition, we used the proposed model to estimate the status of
COVID-19 active cases in the next month.
Although almost the same preventive measures have been taken in almost all areas infected with
COVID-19, in some areas the death or recovery of patients is very different. The prediction of recovery or
death of patients affects the decision of the authorities given the burden of the disease on people's anxiety and
the economy. This study almost comprehensively analyzes COVID-19in terms of: i) predicting the final
status of patients (recovery or death) and ii) the effect of the longitude and latitude of the infected areas on
the final status of active cases separately for each region and continent. Our results indicate that the model
might be generalized in similar pandemics considering that the proposed model does not depend on the
clinical characteristics of the disease and predicts the condition of patients based on the disease pattern in a
region and other infected areas around the world.
The importance of designing such models can be considered for several reasons: i) at the beginning
of epidemics, the outbreak does not occur simultaneously all over the world so the information about the
outbreak is available for all regions. Therefore, if there are models that can predict the conditions of an area
in the coming weeks based on the disease behavior pattern in neighboring areas, preventive measures can be
taken to prevent the spread of the disease more effectively and ii) due to the unknown nature of the disease
and the factors affecting it, the existence of models that can predict the final status of patients based on non-
clinical information can help administrators to manage the allocation of resources and medical staff in
affected areas, and thus increasing the quality of services and reducing the community anxiety. The high
accuracy of the proposed model in predicting the recovery and death in the next 2-week indicates that the
proposed model can be used in the event of other pandemics in the future and can guide planning and
resource allocation for prevention, treatment, and palliative care.
REFERENCES
[1] J. Nkengasong, “China’s response to a novel coronavirus stands in stark contrast to the 2002 SARS outbreak response,” Nature
Medicine, vol. 26, no. 3, pp. 310–311, Jan. 2020, doi: 10.1038/s41591-020-0771-1.

3115
[2] B. Saman, M. M. A. Eid, and M. M. Eid, “Recently employed engineering techniques to reduce the spread of COVID-19 (corona
virus disease 2019): a review study,” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 22,
no. 1, pp. 277–286, Apr. 2021, doi: 10.11591/ijeecs.v22.i1.pp277-286.
[3] K. Roosa et al., “Real-time forecasts of the COVID-19 epidemic in China from February 5th to February 24th, 2020,” Infectious
Disease Modelling, vol. 5, pp. 256–263, 2020, doi: 10.1016/j.idm.2020.02.002.
[4] Eurosurveillance Editorial Team, “Note from the editors: world Health Organization declares novel coronavirus (2019-nCoV)
sixth public health emergency of international concern,” Eurosurveillance, vol. 25, no. 5, Feb. 2020, doi: 10.2807/1560-
7917.ES.2020.25.5.200131e.
[5] WHO, “Weekly epidemiological update on COVID-19,” World Health Organization, 2021.
https://www.who.int/publications/m/item/weekly-epidemiological-update-on-covid-19---20-july-2021 (accessed Mar. 01, 2021).
[6] W. Ji, W. Wang, X. Zhao, J. Zai, and X. Li, “Cross‐species transmission of the newly identified coronavirus 2019‐nCoV,”
Journal of Medical Virology, vol. 92, no. 4, pp. 433–440, Apr. 2020, doi: 10.1002/jmv.25682.
[7] D. Paraskevis, E. G. Kostaki, G. Magiorkinis, G. Panayiotakopoulos, G. Sourvinos, and S. Tsiodras, “Full-genome evolutionary
analysis of the novel corona virus (2019-nCoV) rejects the hypothesis of emergence as a result of a recent recombination event,”
Infection, Genetics and Evolution, vol. 79, Apr. 2020, doi: 10.1016/j.meegid.2020.104212.
[8] C. Huang et al., “Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China,” The Lancet, vol. 395,
no. 10223, pp. 497–506, Feb. 2020, doi: 10.1016/S0140-6736(20)30183-5.
[9] J. Y. Kim et al., “The first case of 2019 novel coronavirus pneumonia imported into korea from wuhan, china: Implication for
infection prevention and control measures,” Journal of Korean Medical Science, vol. 35, no. 5, 2020, doi:
10.3346/jkms.2020.35.e61.
[10] S. Bernard Stoecklin et al., “First cases of coronavirus disease 2019 (COVID-19) in France: surveillance, investigations and
control measures, January 2020,” Eurosurveillance, vol. 25, no. 6, Feb. 2020, doi: 10.2807/1560-7917.ES.2020.25.6.2000094.
[11] M. Giovanetti, D. Benvenuto, S. Angeletti, and M. Ciccozzi, “The first two cases of 2019-nCoV in Italy: Where they come
from?,” Journal of Medical Virology, vol. 92, no. 5, pp. 518–521, Feb. 2020, doi: 10.1002/jmv.25699.
[12] V. M. Corman et al., “Detection of 2019 novel coronavirus (2019-nCoV) by real-time RT-PCR,” Eurosurveillance, vol. 25, no. 3,
Jan. 2020, doi: 10.2807/1560-7917.ES.2020.25.3.2000045.
[13] N. Zhang et al., “Recent advances in the detection of respiratory virus infection in humans,” Journal of Medical Virology, vol. 92,
no. 4, pp. 408–417, Apr. 2020, doi: 10.1002/jmv.25674.
[14] Y. Yang, D. Zhang, C. Zhou, H. Huang, and R. Wang, “Value of lung ultrasound for the diagnosis of COVID-19 pneumonia: a
protocol for a systematic review and meta-analysis,” BMJ Open, vol. 10, no. 8, Aug. 2020, doi: 10.1136/bmjopen-2020-039180.
[15] C. K. Wong et al., “Artificial intelligence mobile health platform for early detection of COVID-19 in quarantine subjects using a
wearable biosensor: protocol for a randomised controlled trial,” BMJ Open, vol. 10, no. 7, Jul. 2020, doi: 10.1136/bmjopen-2020-
038555.
[16] H. B. Mehta, S. Ehrhardt, T. J. Moore, J. B. Segal, and G. C. Alexander, “Characteristics of registered clinical trials assessing
treatments for COVID-19: a cross-sectional analysis,” BMJ Open, vol. 10, no. 6, Jun. 2020, doi: 10.1136/bmjopen-2020-039978.
[17] H. Nishiura et al., “The extent of transmission of novel coronavirus in Wuhan, China, 2020,” Journal of Clinical Medicine, vol. 9,
no. 2, Jan. 2020, doi: 10.3390/jcm9020330.
[18] J. B. Long and J. M. Ehrenfeld, “The role of augmented intelligence (AI) in detecting and preventing the spread of novel
coronavirus,” Journal of Medical Systems, vol. 44, no. 3, Mar. 2020, doi: 10.1007/s10916-020-1536-6.
[19] B. McCall, “COVID-19 and artificial intelligence: protecting health-care workers and curbing the spread,” The Lancet Digital
Health, vol. 2, no. 4, pp. 166–167, Apr. 2020, doi: 10.1016/S2589-7500(20)30054-6.
[20] C.-S. Yu et al., “A COVID-19 pandemic artificial intelligence–based system with deep learning forecasting and automatic
statistical data acquisition: development and implementation study,” Journal of Medical Internet Research, vol. 23, no. 5, May
2021, doi: 10.2196/27806.
[21] H. B. Harchegani, A. Farahi, H. M. Shirazi, A. Golabpour, and P. Almasinejad, “Using genetic algorithm for prediction of
information for cache operation in? mobile database,” in 2009 Second International Workshop on Knowledge Discovery and Data
Mining, Jan. 2009, pp. 148–151, doi: 10.1109/WKDD.2009.201.
[22] R. Torshizi et al., “Altered expression of cell cycle regulators in adult T-Cell leukemia/lymphoma patients,” Reports of
biochemistry and molecular biology, vol. 6, no. 1, pp. 88–94, Oct. 2017.
[23] A. Z. M. Kootiani, M. Doostari, A. Golabpour, and M. Broujerdian, “Differential power analysis in the smart card by data
simulation,” in Proceedings - 2008 International Conference on MultiMedia and Information Technology, MMIT 2008, Dec.
2008, pp. 817–821, doi: 10.1109/MMIT.2008.192.
[24] N. E. Budiyanta, L. Wijayanti, W. Widjaja Basuki, H. Tanudjaja, and V. B. Kartadinata, “The development of healthcare mobile
robot for helping medical personnel in dealing with COVID-19 patients,” Indonesian Journal of Electrical Engineering and
Computer Science (IJEECS), vol. 22, no. 3, Jun. 2021, doi: 10.11591/ijeecs.v22.i3.pp1379-1388.
[25] N. A. Hidayat, P. Megantoro, A. Yurianta, A. Sofiah, S. A. Aldhama, and Y. A. Effendi, “The application of instrumentation
system on a contactless robotic triage assistant to detect early transmission on a COVID-19 suspect,” Indonesian Journal of
Electrical Engineering and Computer Science (IJEECS), vol. 22, no. 3, pp. 1334–1344, Jun. 2021, doi:
10.11591/ijeecs.v22.i3.pp1334-1344.
[26] A. Singh, V. Jindal, R. Sandhu, and V. Chang, “A scalable framework for smart COVID surveillance in the workplace using Deep
Neural Networks and cloud computing,” Expert Systems, May 2021, doi: 10.1111/exsy.12704.
[27] I. Ahmed, M. Ahmad, J. J. P. C. Rodrigues, G. Jeon, and S. Din, “A deep learning-based social distance monitoring framework
for COVID-19,” Sustainable Cities and Society, vol. 65, Feb. 2021, doi: 10.1016/j.scs.2020.102571.
[28] H. M. Haglan, A. S. Mahmoud, M. H. Al-Jumaili, and A. J. Aljaaf, “New ideas and framework for combating COVID-19
pandemic using IoT technologies,” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 22, no. 3,
pp. 1565–1572, Jun. 2021, doi: 10.11591/ijeecs.v22.i3.pp1565-1572.
[29] F. Ahouz and A. Golabpour, “Predicting the incidence of COVID-19 using data mining,” BMC Public Health, vol. 21, no. 1, Dec.
2021, doi: 10.1186/s12889-021-11058-3.
[30] D. A. Gomez-Cravioto, R. E. Diaz-Ramos, F. J. Cantu-Ortiz, and H. G. Ceballos, “Data analysis and forecasting of the COVID-
19 spread: a comparison of recurrent neural networks and time series models,” Cognitive Computation, Jun. 2021, doi:
10.1007/s12559-021-09885-y.
[31] C. Kerdvibulvech and L. (Luke) Chen, “The power of augmented reality and artificial intelligence during the covid-19 outbreak,”
in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics), vol. 12424, Springer International Publishing, 2020, pp. 467–476.
[32] A. W. Reza, M. M. Hasan, N. Nowrin, and M. M. Ahmed Shibly, “Pre-trained deep learning models in automatic COVID-19

 ISSN: 2088-8708
3116
diagnosis,” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 22, no. 3, pp. 1540–1547, Jun.
2021, doi: 10.11591/ijeecs.v22.i3.pp1540-1547.
[33] M. Masum, H. Shahriar, H. M. Haddad, and M. S. Alam, “r-LSTM: time series forecasting for COVID-19 confirmed cases with
LSTMbased framework,” in 2020 IEEE International Conference on Big Data (Big Data), Dec. 2020, pp. 1374–1379, doi:
10.1109/BigData50022.2020.9378276.
[34] K. E. ArunKumar, D. V. Kalaga, C. M. S. Kumar, M. Kawaji, and T. M. Brenza, “Forecasting of COVID-19 using deep layer
recurrent neural networks (RNNs) with gated recurrent units (GRUs) and long short-term memory (LSTM) cells,” Chaos, Solitons
and Fractals, vol. 146, May 2021, doi: 10.1016/j.chaos.2021.110861.
[35] S. J. Fong, G. Li, N. Dey, R. Gonzalez-Crespo, and E. Herrera-Viedma, “Finding an accurate early forecasting model from small
dataset: a Case of 2019-nCoV novel coronavirus outbreak,” International Journal of Interactive Multimedia and Artificial
Intelligence, vol. 6, no. 1, 2020, doi: 10.9781/ijimai.2020.02.002.
[36] A. B. H. Fairoza et al., “Coronatracker: world-wide COVID-19 outbreak data analysis and prediction,” Bulletin of the World
Health Organization, vol. 98, no. 3, Mar. 2020, doi: 10.2471/blt.20.255695.
[37] P. Wang, X. Zheng, G. Ai, D. Liu, and B. Zhu, “Time series prediction for the epidemic trends of COVID-19 using the improved
LSTM deep learning method: Case studies in Russia, Peru and Iran,” Chaos, Solitons & Fractals, vol. 140, Nov. 2020, doi:
10.1016/j.chaos.2020.110214.
[38] H. M. Zawbaa et al., “A study of the possible factors affecting COVID‐19 spread, severity and mortality and the effect of social
distancing on these factors: machine learning forecasting model,” International Journal of Clinical Practice, vol. 75, no. 6, Jun.
2021, doi: 10.1111/ijcp.14116.
[39] H. Pham, “Predictive modeling on the number of covid-19 death toll in the United States considering the effects of coronavirus-
related changes and covid-19 recovered cases,” International Journal of Mathematical, Engineering and Management Sciences,
vol. 5, no. 6, pp. 1140–1155, Dec. 2020, doi: 10.33889/IJMEMS.2020.5.6.087.
[40] “Novel coronavirus (COVID-19) cases data,” Johns Hopkins University Center for Systems Science and Engineering, 2020.
[41] R. Krispin, “Coronavirus,” 2020. https://github.com/RamiKrispin/coronavirus (accessed Mar. 01, 2020).
[42] Z. Boussaada, O. Curea, A. Remaci, H. Camblong, and N. Mrabet Bellaaj, “A nonlinear autoregressive exogenous (NARX)
neural network model for the prediction of the daily direct solar radiation,” Energies, vol. 11, no. 3, Mar. 2018, doi:
10.3390/en11030620.
[43] M. F. Møller, “A scaled conjugate gradient algorithm for fast supervised learning,” Neural Networks, vol. 6, no. 4, pp. 525–533,
Jan. 1993, doi: 10.1016/S0893-6080(05)80056-5.
[44] M. T. Hagan and M. B. Menhaj, “Training feedforward networks with the marquardt algorithm,” IEEE Transactions on Neural
Networks, vol. 5, no. 6, pp. 989–993, 1994, doi: 10.1109/72.329697.
[45] C. L. Giles, S. Lawrence, and A. C. Tsoi, “Noisy time series prediction using recurrent neural networks and grammatical
inference,” Machine Learning, vol. 44, no. 1–2, pp. 161–183, 2001, doi: 10.1023/A:1010884214864.
[46] F. Ahouz and A. Golabpour, “Predicting the COVID-19 prevalence rate using data mining,” BMC Public Health, Apr. 2020, doi:
10.21203/rs.3.rs-21247/v1.
[47] M. Adebiyi and O. O. Olugbara, “Binding site identification of COVID-19 main protease 3D structure by homology modeling,”
Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 21, no. 3, pp. 1713–1721, Mar. 2021, doi:
10.11591/ijeecs.v21.i3.pp1713-1721.
[48] K. E. ArunKumar, D. V. Kalaga, C. M. Sai Kumar, G. Chilkoor, M. Kawaji, and T. M. Brenza, “Forecasting the dynamics of
cumulative COVID-19 cases (confirmed, recovered and deaths) for top-16 countries using statistical machine learning models:
auto-regressive integrated moving average (ARIMA) and seasonal auto-regressive integrated moving averag,” Applied Soft
Computing, vol. 103, May 2021, doi: 10.1016/j.asoc.2021.107161.
[49] “Coronavirus (COVID-19) cases and deaths,” World Health Organization, 2020. https://data.humdata.org/dataset/coronavirus-
covid-19-cases-and-deaths (accessed Mar. 01, 2020).
BIOGRAPHIES OF AUTHORS
Peyman Almasinejad received his Bachelor of Software Engineering from the
Islamic Azad University, Iran in 2006 and Master of Software Engineering from Payame
Noor University (PNU), Tehran, Iran in 2009. Then, He worked in a software company.
Currently, he is working as an instructor at Department of Computer Engineering and
Information Technology, Payame Noor University (PNU), Iran. He can be contacted at
email: p.almasinejad@pnu.ac.ir.
Amin Golabpour I received my PhD in medical informatics from Ferdowsi
University of Mashhad, Mashhad, Iran, and now I am a director of Health Informatics
Technology group of School of Allied Medical Sciences department at Shahroud University
of Medical Sciences, Shahrour, Iran. I received two master's degree in software engineering
and artificial intelligence. He can be contacted at email: a.golabpour@shmu.ac.ir.

3117
Fatemeh Ahouz , I received my M.Sc. in artificial intelligence (AI) from Shahid
Chamran University of Ahvaz, Iran. I've been working as a faculty lecturer at Behbahan
Khatam Alanbia University of Technology, Behbahan, Iran since 2012. In addition to
teaching fundamental courses of computer engineering and soft computing, I've been
conducting some research in deep/machine learning and optimization methods especially in
extracting rules from data. I am interested in practical aspects of AI in engineering and
medical applications. Currently my research works are oriented toward medical informatics.
She can be contacted at email: ahouz@bkatu.ac.ir.
Mohammad-Reza Mollakhalili-Meybodi received his M.Sc. degree in
Computer Engineering from Amirkabir University of Technology, Tehran, Iran, in 2004, and
the B.Sc. degree in Computer Engineering from Shahid Beheshti University in 2001. He
joined the faculty of Computer engineering Department at Islamic Azad University (Maybod
Branch) in 2004. He obtained his Ph.D. degree in Computer Engineering in 2014. His
research interests include Communication Networks, Learning Systems, Algorithms, Graph
Theory, Stochastic Optimization, Software Defined Networks and Soft Computing. He can
be contacted at email: Mollakhalili@maybodiau.ac.ir.
Kamal Mirzaie obtained his B.Sc. in Computer Engineering from Iran
University of Science and Technology (IUST) in 2003 and his M.Sc. in Computer
Engineering from University of Isfahan (UI) in 2005. He obtained his Ph.D. In 2011, he is
currently working as an Assistant Professor at the Department of Computer Engineering,
Maybod Branch, Islamic Azad University, and Maybod, Iran. His research interests include
cognitive science, soft computing, medical data mining, parallel processing, image
processing, and pattern recognition. He can be contacted at email:
k.mirzaie@maybodiau.ac.ir.
Ahmad Khosravi his M.S. and Ph.D. of Epidemiology, Center for Health
Related Social and Behavioral Sciences, he is Associate Professor at Shahroud University of
Medical Sciences (Iran). He can be contacted at email: khosravi2000us@yahoo.com.
Marzieh Rohani-Rasaf her M.S. in 2012, and Ph.D. in Epidemiology in 2019
at Shahid Beheshti University of Medical Sciences (Iran). Currently, she is Assistant
Professor and Head of Student Research Committee at Shahroud University of Medical
Sciences (Iran). She can be contacted at email: rohani@shmu.ac.ir.
Azadeh Bastani received the B.Sc. and M.Sc. degrees in Computer Engineering
from Isfahan University, Isfahan, Iran, and Iran University of Science and Technology,
Tehran, Iran, in 2000 and 2003, respectively. She has been IT expert in Kish University from
2004 to 2008. Since 2008, she has been with the Behbahan Khatam Alanbia University of
Technology, Behbahan, Iran, where she is currently Instructor of the Computer Engineering
Department. Her current research interests include digital image processing, machine
learning, and artificial intelligence systems. She can be contacted at email:
bastani@bkatu.ac.ir.

Predicting the status of COVID-19 active cases using a neural network time series

More Related Content

Similar to Predicting the status of COVID-19 active cases using a neural network time series (20)

More from IJECEIAES (20)

Recently uploaded (20)

Predicting the status of COVID-19 active cases using a neural network time series